Abstract
A recently proposed coevolutionary domain, the Linguistic Prediction Game (LPG), has shown evidence of open-ended dynamics and complexity growth. The game involves combinations of cooperative and competitive interactions within an ecosystem model. Two avenues are given for future work: (1) to find whether the results of the game are truly complex, and (2) to put the model to practical use. The MODES Toolbox is a recent set of metrics for examining such systems. After applying the metrics and performing adaptive pruning, we find that genetic material produced by the LPG, while structurally complex, can be largely nonadaptive in a manner similar to spandrels found in evolutionary biology. To further probe the generality of the evolutionary dynamics, we propose the Collision Game, a new domain with similar characteristics. Despite observing much interesting behavior, we ultimately find that complexity growth appears dependent on the domain, substrate, and selection method. Adopting the perspective of cooptimization and directly seeking a measure of utility, we recast the LPG as the Moore Machine Binary Prediction Task (MM-BPT); however, the problem is found to be too difficult for existing heuristics. This leads us to propose a new algorithm, QueMEU, that maintains a reliable learning gradient in spite of coevolutionary pathologies. In addition to increasing utility on the MM-BPT, certain QueMEU configurations yield strong rates of adaptive complexity growth, novelty, and change by the standards of the MODES Toolbox. Experiments further show that QueMEU outperforms recent alternative cooptimization heuristics across a set of challenging established benchmarks, giving evidence of the algorithm's potential to benefit a wide range of practical applications.